Vibrato is a fast implementation of tokenization (or morphological analysis) based on the Viterbi algorithm.
A Python wrapper is also available here.
Wasm Demo (takes a little time to load the model.)
Vibrato is a Rust reimplementation of the fast tokenizer MeCab,
although its implementation has been simplified and optimized for even faster tokenization.
Especially for language resources with a large matrix
(e.g., unidic-cwj-3.1.1 with a matrix of 459 MiB),
Vibrato will run faster thanks to cache-efficient id mappings.
For example, the following figure shows an experimental result of tokenization time with MeCab and its reimplementations. The detailed experimental settings and other results are available on Wiki.
Vibrato supports options for outputting tokenized results identical to MeCab, such as ignoring whitespace.
Vibrato also supports training parameters (or costs) in dictionaries from your corpora. The detailed description can be found here.
This software is implemented in Rust.
First of all, install rustc and cargo following the official instructions.
You can easily get started with Vibrato by downloading a precompiled dictionary. The Releases page distributes several precompiled dictionaries from different resources.
Here, consider to use mecab-ipadic v2.7.0.
(Specify an appropriate Vibrato release tag to VERSION such as v0.5.0.)
$ wget https://github.com/daac-tools/vibrato/releases/download/VERSION/ipadic-mecab-2_7_0.tar.xz
$ tar xf ipadic-mecab-2_7_0.tar.xz
You can also compile or train system dictionaries from your own resources. See the docs for more advanced usage.
To tokenize sentences using the system dictionary, run the following command.
$ echo '本とカレーの街神保町へようこそ。' | cargo run --release -p tokenize -- -i ipadic-mecab-2_7_0/system.dic.zst
The resultant tokens will be output in the Mecab format.
本 名詞,一般,*,*,*,*,本,ホン,ホン
と 助詞,並立助詞,*,*,*,*,と,ト,ト
カレー 名詞,固有名詞,地域,一般,*,*,カレー,カレー,カレー
の 助詞,連体化,*,*,*,*,の,ノ,ノ
街 名詞,一般,*,*,*,*,街,マチ,マチ
神保 名詞,固有名詞,地域,一般,*,*,神保,ジンボウ,ジンボー
町 名詞,接尾,地域,*,*,*,町,マチ,マチ
へ 助詞,格助詞,一般,*,*,*,へ,ヘ,エ
ようこそ 感動詞,*,*,*,*,*,ようこそ,ヨウコソ,ヨーコソ
。 記号,句点,*,*,*,*,。,。,。
EOS
If you want to output tokens separated by spaces, specify -O wakati.
$ echo '本とカレーの街神保町へようこそ。' | cargo run --release -p tokenize -- -i ipadic-mecab-2_7_0/system.dic.zst -O wakati
本 と カレー の 街 神保 町 へ ようこそ 。
The distributed models are compressed in zstd format.
If you want to load these compressed models with the vibrato API,
you must decompress them outside of the API.
// Requires zstd crate or ruzstd crate
let reader = zstd::Decoder::new(File::open("path/to/system.dic.zst")?)?;
let dict = Dictionary::read(reader)?;Vibrato is a reimplementation of the MeCab algorithm, but with the default settings it can produce different tokens from MeCab.
For example, MeCab ignores spaces (more precisely, SPACE defined in char.def) in tokenization.
$ echo "mens second bag" | mecab
mens 名詞,固有名詞,組織,*,*,*,*
second 名詞,一般,*,*,*,*,*
bag 名詞,固有名詞,組織,*,*,*,*
EOS
However, Vibrato handles such spaces as tokens with the default settings.
$ echo 'mens second bag' | cargo run --release -p tokenize -- -i ipadic-mecab-2_7_0/system.dic.zst
mens 名詞,固有名詞,組織,*,*,*,*
記号,空白,*,*,*,*,*
second 名詞,固有名詞,組織,*,*,*,*
記号,空白,*,*,*,*,*
bag 名詞,固有名詞,組織,*,*,*,*
EOS
If you want to obtain the same results as MeCab, specify the arguments -S and -M 24.
$ echo 'mens second bag' | cargo run --release -p tokenize -- -i ipadic-mecab-2_7_0/system.dic.zst -S -M 24
mens 名詞,固有名詞,組織,*,*,*,*
second 名詞,一般,*,*,*,*,*
bag 名詞,固有名詞,組織,*,*,*,*
EOS
-S indicates if spaces are ignored.
-M indicates the maximum grouping length for unknown words.
There are corner cases where tokenization results in different outcomes due to cost tiebreakers. However, this would be not an essential problem.
You can use your user dictionary along with the system dictionary. The user dictionary must be in the CSV format.
<surface>,<left-id>,<right-id>,<cost>,<features...>
The first four columns are always required.
The others (i.e., <features...>) are optional.
For example,
$ cat user.csv
神保町,1293,1293,334,カスタム名詞,ジンボチョウ
本とカレーの街,1293,1293,0,カスタム名詞,ホントカレーノマチ
ようこそ,3,3,-1000,感動詞,ヨーコソ,Welcome,欢迎欢迎,Benvenuto,Willkommen
To use the user dictionary, specify the file with the -u argument.
$ echo '本とカレーの街神保町へようこそ。' | cargo run --release -p tokenize -- -i ipadic-mecab-2_7_0/system.dic.zst -u user.csv
本とカレーの街 カスタム名詞,ホントカレーノマチ
神保町 カスタム名詞,ジンボチョウ
へ 助詞,格助詞,一般,*,*,*,へ,ヘ,エ
ようこそ 感動詞,ヨーコソ,Welcome,欢迎欢迎,Benvenuto,Willkommen
。 記号,句点,*,*,*,*,。,。,。
EOS
The directory docs provides descriptions of more advanced usages such as training or benchmarking.
We have a Slack workspace for developers and users to ask questions and discuss a variety of topics.
- https://daac-tools.slack.com/
- Please get an invitation from here.
Licensed under either of
- Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.
The initial version of this software was developed by LegalOn Technologies, Inc., but not an officially supported LegalOn Technologies product.
See the guidelines.
Technical details of Vibrato are available in the following resources:
- 神田峻介, 赤部晃一, 後藤啓介, 小田悠介. 最小コスト法に基づく形態素解析におけるCPUキャッシュの効率化, 言語処理学会第29回年次大会 (NLP2023).
- 赤部晃一, 神田峻介, 小田悠介. CRFに基づく形態素解析器のスコア計算の分割によるモデルサイズと解析速度の調整, 言語処理学会第29回年次大会 (NLP2023).
- MeCab互換な形態素解析器Vibratoの高速化技法, LegalOn Technologies Engineering Blog (2022-09-20).
